My theory: almost no one really uses AI agents yet. Like outside the SF tech bubble, where everyone is agent-maxxing, every time I talk to friends outside tech, my in-laws in Texas, or friends here in France, “AI” still basically means whatever features on ChatGPT. And the biggest breakthrough for normal people right now is not autonomous research or multi-agent workflows etc. It’s that ChatGPT can finally send an email or create a calendar invite. it shows how early we are. Also, can you imagine the compute crunch if billions of people actually start using agents to do meaningful parts of their jobs?

I agree with Dwarkesh even if it’s a contrarian take. The standard way of thinking is that AI gets cheaper every year. Better chips, better algorithms, more competition. Intelligence becomes abundant, so the price of compute should keep falling (aka intelligence too cheap to meter) His argument is almost the opposite, at least for the next few years. If frontier models become dramatically more economically useful faster than we can manufacture GPUs, then the value of each GPU rises faster than supply. Compute stops being priced by its cost to produce and starts being priced by what it can earn. How much would you pay to to hire geniuses in a data center? Or if one GPU can generate the output of a great software engineer, why would anyone rent it for today’s prices? The really interesting implication is that the frontier labs end up competing on who can afford the most compute. The labs making the most revenue can bid up GPU prices, making it even harder for everyone else to catch up. Basically it’s a super steep power law with only top labs surviving and the rest fighting to create small cheap models with no pricing power. Long term I still expect compute to get cheap. But during this transition, intelligence could become cheaper while the hardware that produces it becomes dramatically more expensive. That’s a pretty counterintuitive idea but a good idea imo.

My increasingly strong opinion is that compute is the biggest moat in AI. Scientific breakthroughs don’t stay secret for long. Papers get published, researchers move, ideas leak, open source catches up. Give it enough time and almost everyone has access to roughly the same algorithms. What doesn’t diffuse nearly as fast is the ability to run 10x bigger training runs, collect 10x more RL experience, serve billions of inference tokens, and iterate faster because you can afford thousands of experiments in parallel. Scaling compute is brutally hard. It’s GPUs, power, substations, networking, cooling, land, permits, supply chains… and hundreds of billions of dollars. You can’t just decide to build that overnight. The more I look at the industry, the more it feels like the frontier is less about having one genius idea and more about who can compound compute the fastest. Scaling laws reward whoever can keep scaling. That’s a very hard moat to replicate imo!

Guys, the world is so compute constrained. I honestly don’t understand how anyone can think otherwise. Almost nobody is seriously using AI today, and we’re already hitting capacity. Everything is a bottleneck: land for data centers, permits, electricity, grid connections, chips, construction, cooling. Demand is growing far faster than supply, and expanding supply is insanely expensive. In this game, the best-positioned companies are the ones with massive net income from non-AI products. They can fund hundreds of billions in CapEx while everyone else has to raise money just to stay in the race… and remember if you are not at the frontier you have no pricing power so you won’t be able to fund your compute need.

When people say data centers use millions of gallons of water, they're describing an old technology. Before, water went into evaporative cooling towers. Warm water pulled heat off the AI chips, then evaporated into the air to shed it. It was effective, but it burned through fresh water continuously, which is where the headline numbers come from today around "data centers use a massive amount of water." The data centers we revealed at Build today don't work that way. The cooling loop is closed. Water is added once during construction and recirculates indefinitely between the servers and the chillers. No evaporation, no fresh-water resupply! Satya put the scale in plain terms: a full year of water use is roughly what a single restaurant uses. Keep in mind that for us, every liter and every watt is an optimization target. The economics and the environment push in the same direction!